developing a global road safety model - unece€¦ · developing a global road safety model george...
TRANSCRIPT
Developing a Global Road Safety Model
George Yannis1, Eleonora Papadimitriou1, Katerina Folla1
Nenad Nikolic2, Eva Molnar2
1Department of Transportation Planning and Engineering, NTUA, Greece2UN Economic Commission for Europe, Sustainable Transport Division, Switzerland
Washington, D.C., 10 January 2018
NHTSA / ITF-IRTAD – Analysis of International Road Safety data
Background
George Yannis, Professor NTU Athens
• Road accidents constitute a major social
problem in modern societies, with road traffic
injuries being estimated as the eighth leading
cause death globally.
• Particularly in low and middle income countries,
road traffic injuries are twice those in high
income countries and still increasing.
• UN Decade of Action: need to strengthen
global and national efforts for casualty
reduction through evidence-based approaches.
Objective
George Yannis, Professor NTU Athens
• To develop a macroscopic road safety decision
making tool that will assist governments and
decision makers, both in developed and
developing countries, to decide on the most
appropriate road safety policies and measures
in order to achieve tangible results.
• Based on work carried out in the framework of
the “Safe Future Inland Transport Systems
(SafeFITS)” project of the United Nations
Economic Commission for Europe (UNECE),
financed by the International Road Union (IRU).
Conceptual Framework
George Yannis, Professor NTU Athens
Based on the five pillars of WHO Global Plan of
Action (WHO, 2011) and an improved version of
the SUNflower pyramid (2002):
SafeFITS layers
1. Economy and Management
2. Transport Demand and Exposure
3. Road Safety Measures
4. Road Safety Performance Indicators
5. Fatalities and Injuries
SafeFITS pillars
1. Road Safety Management
2. Road Infrastructure
3. Vehicle
4. User
5. Post-Crash Services
Architecture of the Database
George Yannis, Professor NTU Athens
• Data from the five layers and the five pillars
• International databases explored: WHO, UN,
IRF, OECD, etc.
• Data for 130 countries with population
higher than 2,8 million inhabitants
• Data refer to 2013 or latest available year
Economy and Management
George Yannis, Professor NTU Athens
Demographic and Economic Characteristics• Population (World Bank Database)
• Area (World Bank Database)
• GNI per capita in US dollars (World Bank Database)
• Projected GDP per capita for 2015-2030 in 2010 US dollars
(ERS International Macroeconomic Data Set)
Road Safety Management Indicators (WHO)• Existence of RS lead agency
• The lead agency is funded
• Existence of national RS strategy
• The RS strategy is funded
• Existence of RS fatality targets
Transport Demand and Exposure
George Yannis, Professor NTU Athens
Roads
• Road network density (IRF)
• Percentage of motorways (IRF)
• Percentage of paved roads (IRF, CIA)
Vehicles (IRF)
• Number of vehicles in use in total and by type
of vehicle
Traffic (IRF)
• Traffic Volume
• Inland surface passengers transport
• Inland surface freight transport
Road Safety Measures (1/2)
George Yannis, Professor NTU Athens
Roads (WHO)• Road safety audits on new roads
• Existence of speed law
• Max speed limits on urban roads (no speed limits; >50 km/h; ≤50 km/h)
• Max speed limits on rural roads (no speed limits; 100-120 km/h;
70-90 km/h; ≤70 km/h)
• Max speed limits on motorways (no speed limits; ≤100 km/h;
100-120 km/h; ≥120 km/h)
Vehicles• Existence of ADR law (UNECE)
• Vehicle standards include seat-belts, electronic stability control,
pedestrian protection (WHO)
• New cars subjected to NCAP (WHO)
Post-crash care (WHO)• Training in emergency medicine for doctors
• Training in emergency training for nurses
Road Safety Measures (2/2)
George Yannis, Professor NTU Athens
Road User (WHO)• Existence of drink-driving law
• Allowed BAC limits (3 separate variables for general
population, young/novice drivers, commercial drivers)
• Existence of national seat-belt law
• The seat-belt law applies to all occupants
• Existence of national child restraint law
• Existence of national helmet law
• The law requires helmet to be fastened
• The helmet law defines specific helmet standards
• Existence of national law on mobile phone use while
driving
• The law applies to hand-held phones
• The law applies to hands-free phones
• Existence of penalty point system
Road Safety Performance Indicators
George Yannis, Professor NTU Athens
Traffic law enforcement (WHO)• Assessment of effectiveness of seat-belt law enforcement
• Assessment of effectiveness of drink-driving law enforcement
• Assessment of effectiveness of speed law enforcement
• Assessment of effectiveness of helmet law enforcement
Road User (WHO)• Seat-belt wearing rates in front seats
• Seat-belt wearing rates in rear seats
• Helmet wearing rates – driver
Post-crash care• Estimated percentage of seriously injured patients transported
by ambulance (WHO)
• Number of hospital beds per population (World Bank
Database)
Fatalities and Injuries
George Yannis, Professor NTU Athens
• Estimated number of road traffic fatalities (WHO)
• Estimated road traffic fatality rates per 100.000
population (WHO)
• Distribution of road traffic fatalities by road user
type (WHO)
• Distribution of road traffic fatalities by gender
(WHO)
• Percentage of road traffic fatalities attributed to
alcohol (WHO)
Database Overview
George Yannis, Professor NTU Athens
• Wherever data for 2013 were not available, the latest data
available were used.
• The missing values of each indicator of the countries were
filled with the mean value of the indicator in their regions.
• The respective information of each variable is properly
represented in the database for the statistical process.
• Data for most variables were available for almost all
countries.
• Low data availability is observed for few variables regarding:
• the restraint use rates
• the percentage of fatalities attributed to alcohol
• the distribution of fatalities by road user type
• transport demand and exposure indicators
Data Analysis Methodology
• Two-step approach of statistical modeling:
• Estimation of composite variables (factor
analysis) in order to take into account as many
indicators as possible of each layer
• Correlating road safety outcomes with
indicators through composite variables by
developing a regression model with explicit
consideration of the time dimension
• Model specificationLog(Fatalities per Population)ti = Ai + Log(Fatalities per
Population)(t-τ)+ Bi * GDPti + Ki * [Economy & Management]ti + Li
* [Transport demand & Exposure]ti + Mi * [Road Safety Measures]ti+ Ni * [RSPI]ti + εi
Where [Composite Variable]
George Yannis, Professor NTU Athens
Calculation of composite variables – Economy and Management
George Yannis, Professor NTU Athens
[Comp_EM] = -0.250 (EM2_lt15yo) + 0.229
(EM3_gt65yo) + 0.228 (EM4_UrbanPop) + 0.224
(EM7_NationalStrategy) + 0.221
(EM8_NationalStrategyFunded) + 0.222
(EM9_FatalityTargets)
Indicator loadings and coefficients on the estimated
factor (composite variable) on Economy and
ManagementComponent
Loadings Score coefficients
EM1_Popdensity ,091 ,029
EM2_lt15yo -,778 -,250
EM3_gt65yo ,714 ,229
EM4_UrbanPop ,709 ,228
EM5_LeadAgency ,284 ,091
EM6_LeadAgencyFunded ,226 ,073
EM7_NationalStrategy ,697 ,224
EM8_NationalStrategyFunded ,626 ,201
EM9_FatalityTargets ,692 ,222
Calculation of composite variables – Transport Demand and Exposure
George Yannis, Professor NTU Athens
[[Comp_TE] = 0.161 (TE1_RoadNetworkDensity) +
0.149 (TE2_Motorways) + 0.238 (TE3_PavedRoads) +
0.272 (TE4_VehiclesPerPop) + 0.267 (TE5_PassCars) -
0.221 (TE7_PTW) - 0.117 (TE10_PassengerFreight)
Indicator loadings and coefficients on the
estimated factor (composite variable) on
Transport Demand and Exposure
Component
Loadings Score coefficients
TE1_RoadNetworkDensity ,497 ,161
TE2_Motorways ,460 ,149
TE3_PavedRoads ,734 ,238
TE4_VehiclesPerPop ,839 ,272
TE5_PassCars ,825 ,267
TE6_VansLorries -,132 -,043
TE7_PTW -,681 -,221
TE8_Vehkm_Total ,269 ,087
TE9_RailRoad ,136 ,044
TE10_PassengerFreight -,360 -,117
Calculation of composite variables - Measures
George Yannis, Professor NTU Athens
[Comp_ME] = 0.069(ME2_ADR) +
0.045(ME4_SpeedLimits_urban) +
0.064(ME6_SpeedLimits_motorways) +
0.088(ME7_VehStand_seatbelts) +
0.091(ME8_VehStand_SeatbeltAnchorages) +
0.092(ME9_VehStand_FrontImpact) +
0.091(ME10_VehStand_SideImpact) +
0.090(ME11_VehStand_ESC) +
0.087(ME12_VehStand_PedProtection) +
0.090(ME13_VehStand_ChildSeats) +
0.068(ME15_BAClimits) + 0.068(ME16_BAClimits_young)
+ 0.065(ME17_BAClimits_commercial) +
0.057(ME19_SeatBeltLaw_all) +
0.063(ME20_ChildRestraintLaw) +
0.034(ME22_HelmetFastened) +
0.038(ME23_HelmetStand) + 0.038(ME24_MobileLaw) +
0.035(ME25_MobileLaw_handheld) +
0.038(ME27_PenaltyPointSyst) +
0.040(ME29_EmergTrain_nurses)
Indicator loadings and coefficients on the estimated factor (composite variable) on MeasuresComponent
Loadings Score coefficients
ME1_RSA ,245 ,025
ME2_ADR ,681 ,069
ME3_SpeedLaw ,229 ,023
ME4_SpeedLimits_urban ,443 ,045
ME5_SpeedLimits_rural ,200 ,020
ME6_SpeedLimits_motorways ,634 ,064
ME7_VehStand_seatbelts ,877 ,088
ME8_VehStand_SeatbeltAnchorages ,906 ,091
ME9_VehStand_FrontImpact ,908 ,092
ME10_VehStand_SideImpact ,904 ,091
ME11_VehStand_ESC ,891 ,090
ME12_VehStand_PedProtection ,862 ,087
ME13_VehStand_ChildSeats ,896 ,090
ME14_DrinkDrivingLaw ,126 ,013
ME15_BAClimits ,670 ,068
ME16_BAClimits_young ,670 ,068
ME17_BAClimits_commercial ,645 ,065
ME18_SeatBeltLaw ,297 ,030
ME19_SeatBeltLaw_all ,570 ,057
ME20_ChildRestraintLaw ,628 ,063
ME21_HelmetLaw ,236 ,024
ME22_HelmetFastened ,334 ,034
ME23_HelmetStand ,379 ,038
ME24_MobileLaw ,375 ,038
ME25_MobileLaw_handheld ,350 ,035
ME26_MobileLaw_handsfree -,295 -,030
ME27_PenaltyPointSyst ,378 ,038
ME28_EmergTrain_doctors ,178 ,018
ME29_EmergTrain_nurses ,399 ,040
Calculation of composite variables - SPIs
George Yannis, Professor NTU Athens
[Comp_PI] = 0.144 (PI1_SeatBeltLaw_enf) + 0.155
(PI2_DrinkDrivingLaw_enf) + 0.152
(PI3_SpeedLaw_enf)+ 0.160 (PI4_HelmetLaw_enf)
+ 0.155 (PI5_SeatBelt_rates_front) + 0.146
(PI6_SeatBelt_rates_rear) + 0.150
(PI7_Helmet_rates_driver)+ 0.127
(PI8_SI_ambulance) + 0.116 (PI9_HospitalBeds)
Indicator loadings and coefficients on the
estimated factor (composite variable) on SPIs
Component
Loadings Score coefficients
PI1_SeatBeltLaw_enf ,756 ,144
PI2_DrinkDrivingLaw_enf ,812 ,155
PI3_SpeedLaw_enf ,795 ,152
PI4_HelmetLaw_enf ,837 ,160
PI5_SeatBelt_rates_front ,811 ,155
PI6_SeatBelt_rates_rear ,766 ,146
PI7_Helmet_rates_driver ,784 ,150
PI8_SI_ambulance ,667 ,127
PI9_HospitalBeds ,607 ,116
Final Statistical Model
The optimal performing model for the
purposes of SafeFITS
• Dependent variable is the logarithm of the
fatality rate per population for 2013
• The main explanatory variables are the
respective logarithm of fatality rate in 2010
and the respective logarithm of GNI per
capita for 2013
• Four composite variables: the economy &
management, the transport demand and
exposure, the measures, and the SPIs
George Yannis, Professor NTU Athens
Parameter BStd.
Error
95% Confidence Interval Hypothesis Test
Lower UpperWald Chi-
Squaredf p-value
(Intercept) 1,694 ,2737 1,157 2,230 38,291 1 <,001
Comp_ME -,135 ,0646 -,261 -,008 4,358 1 ,037
Comp_TE -,007 ,0028 -,013 -,002 7,230 1 ,007
Comp_PI -,007 ,0030 -,013 -,001 5,652 1 ,017
Comp_EM ,007 ,0051 -,003 ,017 2,009 1 ,156
LNFestim_2010 ,769 ,0462 ,678 ,859 276,322 1 <,001
LNGNI_2013 -,091 ,0314 -,153 -,030 8,402 1 ,004
(Scale) ,038
Likelihood Ratio 1379,00
df 6
p-value <,001
Statistical Model AssessmentIn order to assess the model, a comparison of the observed and the predicted values was carried out:
• The mean absolute prediction error is estimated at 2.7 fatalities per population, whereas the mean
percentage prediction error is estimated at 15% of the observed value.
• The model is of very satisfactory performance as regards the good performing countries (low
fatality rate) and of quite satisfactory performance as regards the medium performing countries.
George Yannis, Professor NTU Athens
0,0
5,0
10,0
15,0
20,0
25,0
30,0
35,0
40,0
0,00 10,00 20,00 30,00 40,00
Pre
dic
ted F
ata
lity
Rate
2013
Observed Fatality Rate 2013
Statistical Model ValidationIn order to validate the model, a cross-validation was carried out with two subsets:
• 80% of the sample was used to develop (fit) the model, and then the model was implemented
to predict the fatality rate for 2013 of the 20% of the sample not used
• 70% of the sample was used to develop (fit) the model, and then the model was implemented
to predict the fatality rate for 2013 of the 30% of the sample not used
George Yannis, Professor NTU Athens
0,0
5,0
10,0
15,0
20,0
25,0
30,0
35,0
40,0
0,00 5,00 10,00 15,00 20,00 25,00 30,00 35,00 40,00
Pre
dic
ted F
ata
lity
Ra
te 2
01
3Observed Fatality Rate 2013
Validation on 30% of the sample
0,0
5,0
10,0
15,0
20,0
25,0
30,0
35,0
40,0
0,00 5,00 10,00 15,00 20,00 25,00 30,00 35,00 40,00
Pre
dic
ted F
ata
lity
Ra
te 2
01
3
Observed Fatality Rate 2013
Validation on 20% of the sample
SafeFITS Model Demonstration
The overall model implementation
includes 3 distinct steps:
• Step 1 – Countries Benchmark
• Step 2 – Forecast with no new
interventions
• Step 3 – Forecast with
interventions
George Yannis, Professor NTU Athens
Model limitations and future improvements• The SafeFITS model was developed on the basis of the most recent and
good quality data available internationally, and by means of rigorous
statistical methods. However, data and analysis methods always have
some limitations.
• Data are primarily directed at vehicle occupants and thus, effects on
road safety outcomes of VRUs may not be captured.
• The effects of interventions may not reflect the unique contribution of
each separate intervention. It is strongly recommended to test
combinations of “similar” interventions (e.g. several vehicle standards,
several types of enforcement or safety equipment use rates etc.)
• The factor analysis procedure does not assume or indicate that a direct
causal relationship exists.
• The calibration with new data will be the ultimate way to fully assess the
performance of the model.
George Yannis, Professor NTU Athens
Benefits for the Policy Makers
• The first global road safety model to be used for policy support
• Global assessments (i.e. monitoring the global progress
towards the UN road safety targets)
• Individual country assessments of various policy scenarios
• A framework which enhances the understanding of road safety
causalities, as well as of the related difficulties.
• Full exploitation of the currently available global data, and use
of rigorous analysis techniques, to serve key purposes in road
safety policy analysis: benchmarking, forecasting and
intervention testing.
• An important step for monitoring, evidence-base and systems
approach to be integrated in decision-making.
George Yannis, Professor NTU Athens
Developing a Global Road Safety Model
George Yannis1, Eleonora Papadimitriou1, Katerina Folla1
Nenad Nikolic2, Eva Molnar2
1Department of Transportation Planning and Engineering, NTUA, Greece2UN Economic Commission for Europe, Sustainable Transport Division, Switzerland
Washington, D.C., 10 January 2018
NHTSA / ITF-IRTAD – Analysis of International Road Safety data